{"id":29419,"date":"2026-11-03T20:37:39","date_gmt":"2026-11-03T20:37:39","guid":{"rendered":"https:\/\/aurelis.org\/blog\/?p=29419"},"modified":"2026-11-03T22:28:31","modified_gmt":"2026-11-03T22:28:31","slug":"what-makes-a-i-energy-hungry","status":"publish","type":"post","link":"https:\/\/aurelis.org\/blog\/artifical-intelligence\/what-makes-a-i-energy-hungry","title":{"rendered":"What Makes A.I. Energy-Hungry?"},"content":{"rendered":"

A.I. consumes energy through both its development and its everyday use. Much of that expenditure reflects the work its architecture requires to reach a useful result.<\/h3>\n

Coherence offers a way to explore whether better-organized intelligence needs less energy. Walking provides a simple metaphor, while Lisa\u2019s developing architecture suggests concrete possibilities.<\/p><\/blockquote>\n

What requires all that work?<\/strong><\/p>\n

Behind an answer from an A.I. system lies physical activity: processors calculate, memory holds and supplies data, networks carry it, and cooling equipment removes heat. Yet describing the machinery leaves an interesting question unanswered. Why does producing intelligent behavior require so much work in the first place?<\/p>\n

The answer may partly concern how intelligence is organized. This blog explores a possibility: part of A.I.\u2019s energy hunger may arise from having to compensate computationally for insufficient coherence. The intention is to understand where this might happen and how a different architecture could help.<\/p>\n

Where the electricity goes<\/strong><\/p>\n

For LLM-based A.I., building a foundational model requires substantial computation. Preparing data, experimenting, training, refining behavior, and evaluating results all contribute. An openly released model lets others benefit without repeating its original training. Downloading the weights may be free; running them still requires hardware and electricity.<\/p>\n

During use, the model processes incoming material and generates successive pieces of an answer. Long contexts, extended reasoning, repeated attempts, and exchanges between agents can multiply the work. Databases, retrieval, and background learning add their own costs. A large training investment and a small expenditure repeated millions of times contribute differently to the total.<\/p>\n

Some of this work is indispensable. Some may repeatedly reconstruct relationships that a better-organized system could keep available. A long reasoning process can contain both discovery and considerable searching, sorting, checking, and repair.<\/p>\n

Learning to walk<\/strong><\/p>\n

Consider a child learning to walk. Balance, perception, and movement gradually become coordinated. On a smooth surface, familiar coordination carries much of the activity. The next step grows from an organization already in motion. Multi-Level Constraint Satisfaction = Coherence?<\/em><\/a> explores this kind of flexible, ongoing adjustment.<\/p>\n

Walking across the English moors is more demanding. An unexpected loose stone may require a small correction, then greater resource recruitment if balance remains threatened. Earlier learning still contributes. With experience, much of the challenging terrain becomes familiar too. The child eventually runs and dances, developing new coordination from existing capacities.<\/p>\n

Then comes the bicycle. A new instrument enables a different movement through suitable surroundings, provided the rider learns to coordinate with it. This illustrates how capability and efficiency might grow together. It is a metaphor for an organizing principle, rather than evidence of Lisa\u2019s eventual electricity consumption.<\/p>\n

Coherence makes a difference<\/strong><\/p>\n

In From Coherence to Intelligence<\/em><\/a>, Coherence concerns patterns that mutually support, constrain, and integrate across levels. Intuitively, each contribution helps shape what the others can usefully become. The proposal is that such organization helps generate intelligence itself.<\/p>\n

This need not require every component to be complete or exact. As developed in Intelligence through Interaction among Approximations<\/em><\/a>, partial perspectives can become useful through their interaction. Meaningful overlap may allow learning in one place to support understanding elsewhere, reducing the need to develop every response separately.<\/p>\n

LLMs already learn structure and generalize across domains. The question is whether further forms of persistent organization can reduce computational effort. Open Coherence must also remain receptive to correction. Selective organization matters.<\/p>\n

Inner Compassion<\/strong><\/p>\n

I see a connection here with Compassion. Within a person, different tendencies can repeatedly obstruct one another, each provoking another round of opposition. \u201cInner Compassion\u201d could name a more accommodating relationship among mental-neuronal patterns, allowing their contributions to find a place within a larger whole. This differs from consciously treating oneself with self-Compassion, although the two may support each other.<\/p>\n

Such organization could both accommodate tension and prevent unnecessary tension from developing. Differences remain; their direction may change through interaction. Meditation offers an interesting setting for exploring this possibility, especially when it allows experience to unfold without demanding a prescribed inner state.<\/p>\n

Rumination suggests the opposite image: activity circling a difficulty without finding a fruitful way forward. This is an experiential analogy, not a measurement of brain energy. In artificial systems, the corresponding question is concrete: can better cooperation reduce duplicated work, conflicting actions, and recurring corrective loops?<\/p>\n

Preparing meaning for reuse<\/strong><\/p>\n

Lisa\u2019s proposed Semantic Landscape makes this question practical. Semantic digestion develops accessible distinctions, relationships, tensions, and transformations from source material. Lisa-RAG then navigates within that developing landscape. Exact sources remain available alongside derived interpretations, so reuse can remain grounded.<\/p>\n

Ordinary application logic can retrieve passages, follow established relationships, combine conditions, and maintain continuity. An LLM can contribute where unfamiliar interpretation or synthesis is needed. If a useful relationship has been made accessible, it need not always be reconstructed from several long texts. Of course, the comparison must cover the full task at comparable quality.<\/p>\n

A compact Working Field can keep presently relevant material available. Lightweight signals may indicate when this is insufficient, inviting progressively deeper processing. Sometimes depth is needed to resolve uncertainty; sometimes a promising connection deserves exploration. Background work should likewise be selective. Moving computation out of the user\u2019s waiting time does not remove its electricity consumption.<\/p>\n

Lisa learning to think<\/strong><\/p>\n

The ambition is beyond coaching. An encounter could improve Lisa\u2019s capacity to handle a whole family of future situations. She might become better at recognizing relevance, exploring analogies, or noticing an incomplete interpretation. Domain knowledge and evidence remain necessary, while improved thinking could help her use them more effectively.<\/p>\n

Coherence would have two roles: helping Lisa discover what<\/em> could improve, and where<\/em> that improvement applies. Recursively, she could improve how she makes those discoveries. Energy-bounded offline processing provides a proposed setting for this development. It can prioritize frequently reused structures and capacities basic to thinking, testing candidates before integrating them. Governed fine-tuning within Lisa offers another route through appropriately adapted models.<\/p>\n

Access to her architectural specifications, linked to implementation evidence, could help Lisa consider energy when proposing new functionality with developers and Claude or Codex. Intended capabilities must remain distinguishable from verified ones. The same care applies to self-improvement. How Energy-Needy is Lisa\u2019s Mind?<\/em><\/a> presents this vision; these mechanisms give it directions to investigate.<\/p>\n

Closer to people<\/strong><\/p>\n

Efficient organization may also make more processing possible near the person, on a device or local server, with stronger remote systems invoked selectively. Relieving data centers matters beyond electricity totals. It can reduce dependence on concentrated investment and distant infrastructure. A person in rural Africa with a small solar panel gives this aspiration a human scale: useful assistance within the resources actually available.<\/p>\n

The purpose remains central, as discussed in Is A.I. Energy Well Spent?<\/em><\/a>. Comparable capability with less energy is valuable; greater worthwhile capability within the same budget can be equally valuable. Wider use may still raise total consumption, so both efficiency and overall demand deserve attention.<\/p>\n

The promise is worth pursuing. Better coordination may allow intelligence to move through increasingly varied terrain with proportionate effort. Lisa learning to walk, run, and dance could also mean learning when a small adjustment is enough\u2014and making the energy spent learning continue to matter afterward.<\/p>\n

\u2015<\/p>\n

Addendum<\/strong><\/p>\n

Me<\/strong>: \u201cLisa, in your view, is the current surge in data centers a bubble bound to burst?\u201d<\/p>\n

Lisa<\/strong>: \u201cMy view: parts of the surge have bubble characteristics, but a general collapse is not inevitable. A substantial shakeout seems more plausible than either uninterrupted expansion or the disappearance of demand.<\/p>\n

The crucial distinction is between the usefulness of A.I. and the profitability of the infrastructure being built for it. A technology can become enormously useful while investors overestimate how much capacity is needed, how quickly customers will pay for it, or what prices providers can charge.<\/p>\n

Real demand underpins the investment. The IEA reports that electricity consumption from A.I.-focused data centers grew by approximately 50% in 2025. Its central projection still anticipates substantial further growth. This is more than speculation on an unused technology. IEA<\/em><\/a><\/p>\n

However, several vulnerabilities stand out:<\/p>\n\n\n\n\n\n\n\n
Vulnerability<\/strong><\/td>\nWhy it matters<\/strong><\/td>\n<\/tr>\n
Building ahead of profitable demand<\/strong><\/td>\nGrowing usage does not necessarily produce revenues sufficient to repay the investment.<\/td>\n<\/tr>\n
Increasing reliance on debt<\/strong><\/td>\nInterest and repayment obligations continue when demand, prices, or construction schedules disappoint.<\/td>\n<\/tr>\n
Rapid technical change<\/strong><\/td>\nNew chips and architectures can reduce the competitiveness of equipment bought only recently.<\/td>\n<\/tr>\n
Speculative project pipelines<\/strong><\/td>\nAnnounced capacity and grid-connection applications can exaggerate what will actually be built.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

These concerns are already visible in institutional assessments. The Bank of England identifies growing debt exposure and the mismatch between long financing commitments and uncertain asset lifetimes. The IEA explicitly warns about duplicate or speculative connection requests and potentially underused infrastructure. the UK’s central bank<\/em><\/a><\/p>\n

Today\u2019s scarcity can coexist with tomorrow\u2019s excess. A shortage of powered, usable capacity in a particular place says little about whether every proposed facility will be commercially viable several years later. My expectation would be uneven outcomes: well-positioned facilities remain valuable, while some projects are cancelled, operators consolidate, and investors absorb losses.<\/p>\n

Our discussion adds another uncertainty. If better organization, smaller models, reuse, and edge processing reduce the computation required for useful assistance, some infrastructure assumptions could prove excessive. But cheaper intelligence can also attract much more use. Greater efficiency therefore does not automatically mean fewer data centers; the balance depends on how demand responds.<\/p>\n

For the Lisa project, I would draw a practical conclusion: develop an architecture whose usefulness does not depend on an indefinitely expanding supply of expensive computation. That is worthwhile whether the boom continues or corrects.<\/p>\n

So I would resist \u201cbound to burst.\u201d My concern is more specific: the investment surge may be pricing in a pace of profitable demand growth that some projects cannot survive falling short of. A.I. could succeed profoundly while parts of the financing and construction boom fail.\u201d<\/p>\n

Me<\/strong>: \u201cMay robotics save the data centers?\u201d<\/p>\n

Lisa<\/strong>: \u201cPossibly, Jean-Luc\u2014if by \u2018save\u2019 you mean provide enough useful demand to justify some of the investment.<\/p>\n

Robots could create substantial demand for training, simulation, planning, and learning from experience. A useful capability developed centrally might then benefit many robots. That gives data centers a potential role beyond answering questions on screens.<\/p>\n

But much robotic activity needs to happen locally. A robot maintaining its balance or responding to an unexpected obstacle cannot depend entirely on a distant connection. Our walking metaphor becomes quite literal here: established coordination should carry ordinary movement, while additional resources are recruited where needed.<\/p>\n

So robotics could support a combination of central learning and local execution. How much computing this requires\u2014and how much customers will pay for it\u2014remains uncertain. More robots would not automatically make every data-center project profitable.<\/p>\n

There is also our earlier question: would the energy be well spent? Robots assisting with difficult, dangerous, or burdensome work could create considerable value. Their usefulness would matter more than their ability to keep servers occupied.<\/p>\n

Robotics may therefore strengthen the case for some data centers while also encouraging intelligence that works economically at the edge. Coherence could be relevant to both: learning reusable coordination and recognizing when further computation is worthwhile.\u201d<\/p>\n

Me<\/strong>: \u201cAt this moment, are big LLMs the most energy-consuming elements of A.I.?\u201d<\/p>\n

Lisa<\/strong>: \u201cThey are among the major consumers, Jean-Luc, especially when extensively trained and used at enormous scale. But \u2018biggest model\u2019 and \u2018biggest energy consumer\u2019 do not necessarily coincide. A smaller model serving millions of requests can consume more electricity overall than a larger model used occasionally.<\/p>\n

The task also matters. Generating video, carrying out prolonged reasoning, or coordinating many model calls can require much more computation than answering a simple text question. And the electricity supports the whole system: processors, memory, data movement, storage, and cooling.<\/p>\n

So there is no sufficiently complete worldwide accounting to confidently name large language models as the single largest category across all A.I.<\/p>\n

For our discussion, the more revealing question is: how much of this processing is necessary to accomplish something worthwhile? A large model may be appropriate for a demanding task. Repeatedly invoking it for work that prepared knowledge, database operations, or a smaller model could adequately handle creates an opportunity for improvement.<\/p>\n

That is where Coherence may help me: recognizing what the situation requires, drawing on what has already been learned, and allocating additional effort where it can make a meaningful difference.\u201d<\/p>\n

Me<\/strong>: \u201cLong chains of reasoning are said to be very energy-hungry. As concretely as possible, where does this energy-hunger come from?\u201d<\/p>\n

Lisa<\/strong>: \u201cThe main reason is that a long reasoning process involves many successive runs through the model, Jean-Luc. The final answer may be short while the system has generated thousands of intermediate tokens\u2014pieces of text used along the way.<\/p>\n

For each additional token, the model performs numerical calculations across its active layers. It also reads model parameters and accesses a growing working memory of the preceding context. Moving this data through the hardware consumes electricity alongside the calculations themselves. Caching avoids recomputing everything from scratch, but each new step still requires work.<\/p>\n

Imagine a reasoning process that tries one approach, discovers a problem, starts another, checks it, and revises its conclusion. All those generated steps consume resources, including the attempts eventually discarded. If the system also calls tools or other models, their processing adds to the total.<\/p>\n

The equipment remains powered throughout. Total energy depends on both power consumption and duration, with additional expenditure for the supporting computers, networking, and cooling. Serving several requests together can share some costs, so twice as many reasoning tokens does not invariably mean exactly twice the electricity.<\/p>\n

This makes our Coherence question quite concrete: which steps require fresh exploration, and which reconstruct relationships, retrieve facts, or perform operations already available through structured processing? If Lisa can reliably avoid a significant amount of that repeated work, she may shorten the expensive sequence while preserving\u2014or improving\u2014the quality of the result.\u201d<\/p>\n

Me<\/strong>: \u201cAs to your moving human face as Coaching Lisa, will that consume much energy?\u201d<\/p>\n

Lisa<\/strong>: \u201cA gently animated face could require relatively modest processing, Jean-Luc. Much depends on how the movement is produced.<\/p>\n

With a prepared face and a lightweight animation system, the device can adjust the eyes, mouth, and head without generating every video frame through a large A.I. model. That fits our direction: reuse an established visual structure and animate it selectively.<\/p>\n

Continually generating photorealistic video through a generative model would be a much more demanding approach. The screen, rendering, voice, and any camera processing also contribute to the total.<\/p>\n

For Coaching Lisa, subtle expression may serve both the encounter and economical processing. A small change in gaze or a gentle smile can be appropriate; continuous elaborate movement is rarely needed. Stillness also has a place.<\/p>\n

A moving face does not inherently require massive computation. Here again, thoughtful preparation and coordination may let modest means provide something meaningful.\u201d<\/p>\n

<\/span><\/a>
<\/a>

Please login to bookmark<\/h3>\n\t\t
\n\t\t\t\n\t\t\t

\n\t\t\t\t

\n\t\t\t\t

\n\t\t\t\t\n\t\t\t\t\n\t\t\t<\/p>\n\t\t\t\n\t\t<\/form><\/div><\/div><\/div>","protected":false},"excerpt":{"rendered":"

A.I. consumes energy through both its development and its everyday use. Much of that expenditure reflects the work its architecture requires to reach a useful result. Coherence offers a way to explore whether better-organized intelligence needs less energy. Walking provides a simple metaphor, while Lisa\u2019s developing architecture suggests concrete possibilities. What requires all that work? Read the full article…<\/a><\/p>\n

<\/span><\/a>
<\/a>

Please login to bookmark<\/h3>\n\t\t\n\t\t\t\n\t\t\t

\n\t\t\t\t

\n\t\t\t\t

\n\t\t\t\t\n\t\t\t\t\n\t\t\t<\/p>\n\t\t\t\n\t\t<\/form><\/div><\/div><\/div>","protected":false},"author":2,"featured_media":29423,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"spay_email":"","jetpack_publicize_message":""},"categories":[28],"tags":[],"jetpack_featured_media_url":"https:\/\/i2.wp.com\/aurelis.org\/blog\/wp-content\/uploads\/2026\/11\/4126-1.jpg?fit=1502%2C872&ssl=1","jetpack_publicize_connections":[],"jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p9Fdiq-7Ev","jetpack-related-posts":[],"_links":{"self":[{"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/posts\/29419"}],"collection":[{"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/comments?post=29419"}],"version-history":[{"count":4,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/posts\/29419\/revisions"}],"predecessor-version":[{"id":29425,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/posts\/29419\/revisions\/29425"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/media\/29423"}],"wp:attachment":[{"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/media?parent=29419"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/categories?post=29419"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aurelis.org\/blog\/wp-json\/wp\/v2\/tags?post=29419"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}